【作者】 张锐；

【导师】 尹东；

【作者基本信息】 中国科学技术大学 ， 信息与通信工程， 2020， 硕士

【摘要】 近年来,随着人工智能技术的不断发展,智能安防被广泛的应用到实际生活中,为平安城市提供保证,保障人民的安全。如今,城市中各个重要位置均安装有摄像头,一个庞大的天网时刻预防、打击犯罪活动。依靠视频监控对行人轨迹进行描述和对特定人员进行检索成为重要的任务。这其中步态识别便是关键技术之一。步态识别通过人的走路姿态和体型进行身份识别,是重要且稳定的生物特征识别技术,该研究有极高的应用价值。基于深度学习的步态识别研究取得了较大的进展,同时多个公共步态数据集的建立也为算法的评估提供平台。然而,当前步态识别的研究工作仍存在一定问题。首先,主流的步态数据集均在无背景干扰的室内采集,志愿者按固定路径行走进行拍摄。使用背景差分法或帧差法对视频图像进行处理得到行人轮廓图。这种模拟场景下的数据及轮廓提取方法与真实场景存在巨大差异,使得研究与实际应用有较大鸿沟。其次,当前的步态识别工作多采用浅层网络进行研究。尽管浅层网络能关注到局部特征且不易发生过拟合现象,但其有限的学习能力及全局特征的缺失导致识别性能较低。随着计算机视觉通用网络的不断改进,如何使用更深的网络来提升步态识别的准确率成为当务之急。最后,步态识别任务中存在跨视角识别的难点。对于同一行人,轮廓和行走轨迹随拍摄视角变化而改变明显,导致跨视角匹配的准确率相较于同视角匹配降低较大。针对上述步态识别研究中存在的问题,论文做了以下工作:1.设计了一个针对监控场景的步态轮廓图提取方法,并构建监控场景步态识别数据集。首先提出一个对监控场景中环境复杂、光照条件变化有很好鲁棒性的步态轮廓图提取方法。此方法包含检测模块、追踪模块、轮廓提取模块。针对视频图像中行人分辨率占比较小的问题,论文增加检测网络特征图尺寸,调整初始化参数。针对追踪模块提出了一个强特征追踪器。最后结合轮廓提取模块,建立了一个校园监控步态数据集,以评估步态识别算法的性能,搭建步态识别研究与应用间的桥梁。2.提出了一个多尺度步态识别网络,选取合适的骨架网络,融合步态能量图的全局特征和局部特征,为步态识别任务提取更有辨别力的特征。整个网络结合Arcface损失函数和难样本三元损失函数进行学习。在多损失函数的监督下,训练一个端到端的多尺度识别网络。实验证明,Arcface损失函数配合难样本三元损失函数可以最大程度的提升网络对步态特征的辨别能力。网络中全局特征和局部特征在一定程度上进行互补,融合后的特征有更强的辨别力。3.实现了一种基于生成对抗网络的跨视角步态识别方法。在步态识别任务中,跨视角识别是最具挑战性的任务。同一行人不同视角下拍摄的轮廓图存在明显差异。本文设计了一个对抗网络结构并加入三元身份损失,将不同视角下的步态能量图转化成固定视角下的步态能量图,同时保留行人身份信息。实验证明生成图像的特征可辅助原始图像得到更高准确率的识别结果。更多还原

【Abstract】 In recent years,with the continuous development of artificial intelligence,intelligent security has been widely used in real life.It provides a guarantee for safe cities and ensures the safety of the people.Nowadays,surveillance cameras are installed in every important location in the city to prevent criminal activities.Relying on video surveillance,it is an important task to describe pedestrian trajectories and retrieve specific persons.Among them,gait recognition is one of the key technologies.Gait is an important and stable biometric recognition technology based on human walking posture and body shape.Therefore,the study of gait recognition has extremely high application value.Research on gait recognition based on deep learning has made great progress.The establishment of public gait datasets also provides a platform for algorithm evaluation.However,there are still some problems in recent research.First,gait images are collected indoors without any background interference and volunteers walk along a fixed path.Background subtraction and frame difference method are used to get pedestrian contour.However,the data and contours extraction methods in simulated scene are different from real scene,which makes a big gap between research and application.Secondly,recent gait recognition work mostly uses shallow neural networks for research.The shallow network can pay attention to local features and is not prone to overfitting.But its limited learning ability and lack of global features lead to low recognition performance.With the continuous improvement of computer vision network,how to use a deeper network to improve the accuracy of gait recognition has become an urgent task.Finally,there are difficulties in cross-view gait recognition.The contour and walking trajectory change significantly with the change of shooting angle.It causes a greater reduction in the accuracy of cross-view matching compared to the same-view matching.To solve the problems of gait recognition,this thesis does the following work:1.This thesis designs a gait contour extraction framework for monitoring scene and constructs a gait recognition dataset.It is robust to outdoor environments with complex environments and changing lighting conditions.The method includes detection module,tracking module and contour extraction module.To solve the low resolution of pedestrian in video image,this thesis increases the size of the feature map and adjusts the initialization parameters.A strong feature tracker is proposed for tracking module.Finally,combining the constour extraction module,a campus monitoring gait dataset is established to evaluate the gait recognition algorithm.It also builds a bridge between gait recognition research and application.2.This thesis proposes a multi-scale convolutional neural network for gait recognition.This thesis selects a better backbone,and merges the global and local features of the gait energy image to extract more discriminative features.Network combines Arcface loss function and Hard Triplet loss function for supervised learning.Under the supervision of multiple loss functions,we train an end-to-end gait recognition network.Experiments show that the Arcface loss function combined with Hard Triplet loss function can maximizes the network’s ability to distinguish gait features.Meanwhile,the combination of global features and local features has stronger discrimination ability.3.This thesis implements a cross-view gait recognition gait recognition method based on generative adversarial networks.Among gait recognition tasks,cross-view recognition is the most challenging task.There are obvious differences in the images taken by the same pedestrian under different angles.This thesis designs a generative adversarial network structure and a triplet identity loss to convert the gait energy image from different angles into a fixed angle gait energy image while retaining pedestrian identity information.Experiments show that the features of the generated image can assists the original image to obtain higher accuracy recognition results.更多还原